We are interested in exploring different visual patterns by training machine learning (ML) classifiers on raw and foreground images for MI detection, which has been less studied. In this work, we train machine learning classifiers on raw images (containing background lines) and clear images (containing just foreground/object with background lines removed). Two ECG record datasets containing normal (N) and myocardial infarction (MI) data are analysed via high-level features provided by standard 12-lead ECG signals. Only the limb lead I was cropped from the 12-lead signals to generate the input data. Data augmentation was used for a balanced dataset to prevent overfitting while maintaining the required spatiotemporal invariances for a correct diagnosis. The newly generated ‘clear’ dataset results show that the proposed model achieves high classification performance for the AD, KNN and RF models, with accuracies that are 32.1, 27.3 and 18.5% higher than those of their ‘raw’ counterparts, respectively. These results prove the robustness of the model.